Bayesian model averaging for ligand discovery.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Pharmacology

Background:

• High-throughput screening (HTS) is crucial for identifying drug-like molecules in pharmaceutical research.
• Integrating biological data with molecular properties is a key challenge in drug discovery.
• Existing methods for analyzing large molecular datasets have limitations.

Purpose of the Study:

• To present a novel Bayesian approach using Markov chain Monte Carlo (MCMC) simulations for learning classification trees.
• To develop a new method for pharmacophore and ligand discovery.
• To compare the performance of the Bayesian method against Neural Networks (NN) and Support Vector Machines (SVM).

Main Methods:

• Bayesian analysis of high-dimensional descriptor data.
• Markov chain Monte Carlo (MCMC) simulations for classification tree learning.
• Training and assessment using experimentally determined pyruvate kinase binding affinity data.
• Application to a database of over 3.7 million molecules.

Main Results:

• The Bayesian classification tree algorithm achieved high specificity and sensitivity.
• The method naturally handles test sets with missing data.
• It provides a robust ranking of classified compounds.
• Outperformed Neural Network and Support Vector Machine architectures in key metrics.

Conclusions:

• The developed Bayesian algorithm is a powerful tool for pharmacophore and ligand discovery.
• It offers advantages in classifying compounds and handling missing data.
• This approach can significantly aid in selecting and ranking potential biologically active compounds for drug development.